Skip to content

Speculative Quadtree MRP Decoding

Speculative Quadtree Markov Random Process (MRP) Decoding accelerates autoregressive generation by speculating multiple future tokens concurrently in a hierarchical 2D tree topology.

Quadtree Tree Verification

graph TD
    Root["Current Token"] --> Q1["Branch 1 (Top-Left)"]
    Root --> Q2["Branch 2 (Top-Right)"]
    Root --> Q3["Branch 3 (Bottom-Left)"]
    Root --> Q4["Branch 4 (Bottom-Right)"]
    Q1 --> L1["Leaf 1.1"]
    Q1 --> L2["Leaf 1.2"]
    Q2 --> L3["Leaf 2.1"]
    Q2 --> L4["Leaf 2.2"]
  1. Lightweight Draft Head: Computes 4 candidate branches in parallel using a low-overhead linear projection.
  2. Fused Flash-Tree Attention: The target model evaluates the entire tree candidate set in a single forward pass using custom attention tree-masking.
  3. Acceptance Rate: Achieves an average acceptance rate of 2.8 to 3.4 tokens per forward step, doubling throughput without altering final probability distributions.